Weird results with test on train with imagenet-multiGPU.torch code.

[jb]

Hi folks,

I'm getting up to speed with Torch7 and as an initial learning exercise, I trained up a CNN via the code in https://github.com/soumith/imagenet-multiGPU.torch specifically the model in described alexnet_cudnn.lua on a five class problem. Training converged with +97% top1 accuracy (clearly over trained). I then loaded the model separately and tested on the training data expecting to get nearly perfect results matching the training top1 - but I didn't - in fact its far worse. the confusion matrix for the five classes is below:

      16      11      17     366     235

      83      47       8    1483     563

       3      73       6     221     482

     102      97      31    1681    2835

     913     303     168    8475   13394

I've been banging my head on this for past couple of days - training was using the original code directly without modification, testing code is attached.

greatly appreciate any feedback and/or pointers.

thanks!!

[Sergey Zagoruyko]

you should check that the layout of classes is the same. The mapping output class -> target class/human readable class might be different between your runs.

[jb]

thanks for the suggestion, I checked the class layout and its the same :( 

I'm starting over and training on the imagenet 2012 dataset - hopefully that will match the results in the repository readme and I can go from there.

thanks!!